Workflow Planning

Workflow Planning

When to Use

Use workflow planning when:

• The task spans multiple state transitions (e.g., parse → compute → filter → cluster).
• You are unsure which tool comes next or what prerequisites exist.
• You want to see all possible paths through a domain's tool graph.

Do NOT use for simple single-tool tasks — just search and call directly.

Not all servers expose workflow planning tools. They are only registered when

the server has tools with state annotations (requires/produces). If

plan_{server}_workflow is not listed in the server's available tools, skip

planning and use search_{server}_tools to find what you need directly.

The State Graph

Domain tools declare state transitions:

• requires — states that must exist before this tool can run.
• produces — states this tool generates after successful execution.

These form a directed graph that plan_{server}_workflow can navigate.

Available Modes

Overview — see the full graph

plan_{server}_workflow(mode="overview")

Returns all tools with their state requirements and productions. Use this first

to understand the domain's structure.

Path — get a recommended sequence

plan_{server}_workflow(
    mode="path",
    current_state="{domain}.data_loaded",
    target_state="{domain}.results_exported"
)

Returns an ordered list of tools to call to reach the target state from the

current state. Follow this sequence in your execute_{server}_code calls.

Tool — inspect a specific tool's context

plan_{server}_workflow(mode="tool", tool_name="process_data")

Returns the tool's prerequisites, what it produces, and where it fits in the graph.

Workflow with Skills

For common multi-step workflows, load a pre-built skill instead of planning manually:

• search_{server}_tools(query="...", category="skills") — find relevant skills.
• load_{server}_skill(skill_name="...") — load the skill's step-by-step instructions.
• Follow the loaded instructions, which reference specific tools in the correct order.

Skills are preferred over manual planning when they exist for your task — they

encode domain expertise and edge-case handling that raw graph traversal does not.